Joint Modeling of Prediction and Behavioral Patterns for Reliable Recommendation With Implicit Feedback

Implicit feedback in recommender systems contains noisy interactions that do not reflect true user preferences, which diminishes recommendation accuracy. Current methods identify unreliable interactions using single indicators such as loss values or model disagreement, but fail to capture the complex patterns that separate genuine preferences from noise. In this paper, we present a method that estimates interaction reliability by analyzing two key signals: the prediction confidence of the recommender model and the consistency of user behavior across interaction sequences. Our approach uses a two-component two-dimensional Gaussian mixture model to estimate reliability by combining prediction scores with co-occurrence patterns that measure behavioral consistency. Instead of making binary decisions, we implement an adaptive weighting mechanism that allows interactions to influence training according to their estimated reliability. Evaluation on three public datasets (Yelp, Amazon-book, and ML-1M) across three recommendation architectures demonstrated that our method improved recommendation accuracy by up to 38% compared with standard training and 33% over existing denoising approaches. These improvements persisted across various models and evaluation metrics, thereby demonstrating the effectiveness of our reliability estimation approach. Our implementation is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/meruemon/JMPBP</uri>.